Too Much to Lose, or More to Gain? Should Sweden Join the Euro?

Transcription

1 Too Much to Lose, or More to Gain? Should Sweden Join the Euro? J James Reade University of Birmingham Ulrich Volz German Development Institute April 22, 2010 Abstract This paper considers the costs and benefits of Sweden joining the European Economic and Monetary Union (EMU) We pay particular attention to the costs of abandoning the krona in terms of a loss of monetary policy independence For this purpose, we apply a cointegrated VAR framework to examine the degree of monetary independence that the Sveriges Riksbank enjoys Our results suggest that Sweden has in fact relatively little to lose from joining EMU, at least in terms of monetary independence We complement our analysis by looking into other criteria affecting the cost-benefit calculus of monetary integration, which, by and large, support our positive assessment of Swedish EMU membership JEL Classification: E52, E58, F41, F42, C32 Keywords: Swedish EMU membership, Monetary policy independence, European monetary integration, Cointegrated VAR method We would like to thank Marco Barassi, Anindya Banerjee, John Fender and other participants at a seminar at the University of Birmingham for useful comments and suggestions Address for correspondence: Department of Economics, University of Birmingham, JG Smith Building, Birmingham, B15 2TT, UK jjreade@bhamacuk Tel: , Fax: Corresponding Author Address for correspondence: German Development Institute, Tulpenfeld 6, Bonn, Germany ulrichvolz@die-gdide Tel: , Fax:

2 I Introduction The Swedish people in 2003 delivered a resounding rejection of euro membership in a referendum, and the popular perception was that to join the euro would involve relinquishing monetary policy independence to a pan-european body 1 However, was that necessarily true? Sweden would certainly have handed over monetary sovereignty to the European Central Bank (ECB), but did it actually have any monetary independence to lose? If we define monetary independence as the ability of an economy to set its own interest rates exclusive of outside influence, then this is something that can be tested In this paper we ask the timely and relevant question: does Sweden have any operational independence for its monetary policy? The answer to this question has serious implications for the cost-benefit calculus of Swedish EMU membership, as the loss of monetary policy independence is widely regarded as the main cost of entering a monetary union Unlike Denmark and the UK, both of which negotiated an opt-out clause to the third stage of EMU, Sweden is required by EU law to join the common monetary union and adopt the euro as soon as it fulfils the convergence criteria as laid out in the Maastricht treaty However, as Sweden did not join the Exchange Rate Mechanism (ERM) of the European Monetary System prior to the creation of EMU, nor ERM II which it needs to be part of to fulfil the Maastricht criterion on exchange rate stability to join EMU, it has de facto opted to stay outside the euro area 2 Unlike other EU members that are obliged to join the euro, Sweden does not stabilise its currency against the euro (Figure 1) Instead, the Sveriges Riksbank has followed a policy of inflation targeting with a flexible exchange rate since 1992/1993, which in theory should provide it with full monetary autonomy, a condition we seek to examine While Swedish euro membership disappeared from the political agenda and any serious debate about it was more or less muted after the negative outcome of the 2003 referendum, the current global financial crisis has brought back the question of EMU membership Like in other small open European economies 1 Jonung and Vlachos (2007, pp 54 5), who analyse the 2003 rejection of euro membership from a political economy perspective, describe the outcome as follows: No-voters saw the euro as a threat to national independence, Swedish democracy and the Swedish way of life They feared that joining the euro meant that decisions of major importance were taken out of the hands of domestic voters and domestic politicians and transferred to Frankfurt and Brussels to be made by policy-makers that were not democratically accountable according to their view 2 Sweden has fulfilled all other Maastricht criteria See European Commission (2008) 2

3 11 E xc h a n g e R at e Time Figure 1: The krona to euro exchange rate Source: Datastream such as Denmark, Iceland and the Central and Eastern European EU members that have not joined the euro yet the costs and benefits of remaining outside the euro area are now being reassessed in Sweden Apparently, public opinion in Sweden, as in other euro area outsider countries (with the notable exception of the UK), has been recently turning in favour of joining the euro 3 By examining the degree of monetary policy independence that the Sveriges Riksbank has enjoyed over the past years, we address what is commonly regarded as the main cost of entering a monetary union; if Sweden de facto enjoys little or no monetary policy independence, then the costs of abandoning the krona and replacing it with the euro were much smaller than opponents of Sweden s euro membership make believe While the question of Sweden s monetary independence is at the heart of our analysis, this paper also enters into the broader nature of Sweden s economic relationship with the eurozone to help frame the question of euro membership The paper is structured as follows In Section II we briefly review the previous literature on Swedish EMU membership In Section III we introduce the econometric methods employed in this paper, followed by an introduction of the data in Section IV Section V reports the results of the econometric analysis In Section VI we provide a broader analysis of the costs and benefits of Swedish euro membership and consider possible convergence between Sweden and the eurozone Section VII concludes 3 See 3

4 II Literature Review The economics profession has produced a vast amount of output on the costs and benefits of European monetary integration 4 The amount of research carried out on Swedish EMU membership has been far less extensive; especially after the negative referendum outcome, research has been scant The Swedish debate has been influenced very much by the report on the consequences of Swedish membership in the monetary union that was prepared by the so-called Calmfors Commission (Calmfors et al, 1997) The Calmfors Commission was appointed by the Swedish government to examine the consequences of Swedish membership in the monetary union ahead of the Riksdag s (the Swedish parliament) decision in 1997 on whether or not Sweden should join EMU The Commission argued that entering EMU would bring about small efficiency gains for the Swedish economy due to reduced transaction costs and exchange rate uncertainty on the one hand and stronger competition on the other The Commission contrasted these benefits with potentially adverse effects that large Sweden-specific disturbances could have if these were not responded to adequately by an independent monetary and exchange rate policy While large idiosyncratic shocks would remain the exception rather than the rule, the Commission maintained that an independent monetary policy would be an insurance for the occurrence of such extreme events and that the potential stabilisation policy cost of relinquishing monetary policy independence may be considerable (Calmfors et al, 1997, p 314) It consequently argued that Sweden should not immediately join EMU upon its creation in 1999, but rather at some later stage when the experience with EMU as well as Sweden s outsider status could be evaluated 5 A second government commission published a report in 2002, the year before the referendum (SOU, 2002b, p 1), where it highlighted again that [m]embership in the monetary union will mean a change in the stabilisation policy regime because domestic monetary policy will disappear as an instrument to stabilise the economy Instead, Sweden will participate in a common European monetary and exchange rate policy The opportunity to use interest rate changes to counteract macroeconomic shocks that specifically affect the Swedish economy will then be lost 6 4 For an overview see, for instance, Emerson et al (1992) and De Grauwe (2007) and the works cited therein 5 It is noteworthy that the Riksbank took a stand in favour of EMU membership when asked by the government to respond to the Calmfors report 6 A summary of the report is avalaible in English under SOU (2002a) 4

5 In a recent contribution Söderström (2008) noted that more than a decade after the Calmfors report was released some of its original arguments seem to speak more strongly in favour of Swedish EMU membership, whereas other arguments speak more strongly against Regarding the former, Söderström concludes that the gains of economic integration induced by EMU, such as increases in trade and financial integration, appear to be greater than expected by Calmfors et al 7 Furthermore, Söderström shows that the Swedish business cycle has been closely correlated with the economies of the euro area since the mid-1990s, suggesting that common shocks have been an important driving force of business cycles across Europe Moreover, he points out that there have been no large country-specific disturbances to the Swedish economy, which would have yielded substantial gains from an independent monetary policy Söderström (2008) comes to no clear-cut conclusion about whether or not participation in EMU would be advantageous for Sweden, but suggests that the current financial crisis might deliver interesting insights into the cost-benefit analysis surrounding Swedish membership In this context, a recent paper by Buiter (2008) provides arguments in favour of Swedish EMU entry Although Buiter s analysis is concerned exclusively with the question of the United Kingdom s EMU membership, his financial stability arguments apply also to Sweden maybe even more than to the UK Buiter argues that, without eurozone membership, the UK is more vulnerable to a triple financial crisis, namely a banking, currency and sovereign debt crisis According to Buiter, the UK belongs to a group of countries characterised by the inconsistent quartet : (1) a small country with (2) a large internationally exposed banking sector, (3) a currency that is not a global reserve currency and (4) limited fiscal capacity relative to the possible size of the banking sector solvency gap All characteristics of this inconsistent quartet certainly apply to Sweden as well, exposing it to potentially damaging vulnerabilities 8 As Buiter points out, joining EMU would immediately eliminate the third pillar of the quartet, and maybe even provide some relief as regards the fourth pillar of the quartet by reducing liquidity risk premia 9 7 For recent studies on trade effects of the euro see Baldwin (2006) and Flam and Nordström (2007) In the early 2000s, Rose (2001) estimated that in case of adopting the euro, Sweden s trade with the euro area would rise by over fifty percent 8 The heavy exposure of Swedish banks in Eastern Europe (and particularly in the crisis-hit Baltic states) has led to growing concerns about a sharp rise in loan losses for Sweden s main banks, which has prompted the Riksbank in 2009 to activate a swap agreement with the ECB to ensure it has enough currency reserves to guarantee stability of its banking sector 9 For a detailed theoretical analysis of the costs and benefits of the UK (or any other small open European economy) joining EMU see Buiter (2000) 5

6 In an assessment of the performance of EMU and an analysis of whether Denmark, Sweden and the UK should join the eurozone, Flam et al (2008) argue that given that the Swedish central bank has closely followed the ECB s monetary policy, Sweden s macroeconomic performace would probably have been quite similar to its actual performance if it had joined EMU in 1999 In the assessment of Flam et al, the main difference for Sweden would have been that its nominal effective and therefore real exchange rate would have been more stable, and that trade with the eurozone would have increased substantially Flam et al (2008, p 17) come to the conclusion that [t]he benefit of an independent monetary policy has [] been small during the past ten years, and should hence decrease further with increasing economic integration Other recent work addressing the question of Swedish euro membership either directly or indirectly has been carried out by Pesaran et al (2007) and Ferreira-Lopes (2008) 10 Pesaran et al (2007) use a Global VAR model to estimate effects of a hypothetical entry of the UK and Swedish economies to the euro area For Sweden their model suggests a likely increase in both output and prices, which leads them to conclude that no unambiguous welfare conclusions are possible for Sweden Söderström flags up foreign exchange shocks as a key source of macroeconomic instability for Sweden, to some extent resonating with Buiter, indicating that the exchange rate has to a large extent destabilised rather than stabilised the Swedish economy This suggests that for a small open economy like Sweden, there are additional costs to maintaining a currency that is not a global reserve currency in terms of the shocks that affect it Ferreira-Lopes and Söderström adopt a different empirical approach to that taken in this paper; our approach might be described as econometric-theory-centric economic modelling while their approach would better be described as economic-theory-centric economic modelling Our terminology is merely for convenience; other terms used to describe these schools of modelling are reduced-form and structural-form modelling It is our belief that both approaches have great value to add to debate on economic policy Economic-theory-centric modelling, by modelling at the structural level, explicitly models the preferences of 10 Several other recent contributions concerned with Swedish EMU membership are in fact more about supposedly deficient structures of EMU, such as the Stability and Growth Pact (cf Jespersen, 2004), about a lack of democratic legitimacy of the ECB s monetary policy making (Forder, 2004) or about the ECB s monetary strategy and inflation performance (Vaubel, 2004) than about economic conditions for Swedish euro entry 6

7 economic agents and can enable analysis that is robust to the Lucas (1976) critique, in that it will account for the impact of policy actions on agent preferences Furthermore, such economic-theory-centric modelling also allows counter-factual analysis, by allowing the modeller to simulate a model where a policy is enacted, and one in which it is not This is something an econometric-theory-centric modelling strategy cannot do, because it is based on observational data This does not, however, render econometric-theory-centric modelling useless in light of the Lucas critique If econometric models can be shown to exhibit super exogeneity (Engle et al, 1983), so that parameter estimates in the econometric model are invariant to known historical policy interventions, then the estimates can be seen as reliable for policy analysis If an econometric model, motivated by economic theories, can be found that satisfies the assumptions placed on it (notably Normality of the error terms), and also exhibits super exogeneity, then this constitutes an equally useful tool for economic policy making as an economic-theory-centric model This modelling methodology is perhaps best explained and examined in Hendry (1995) and embodied in the general-to-specific, LSE approach to econometrics The disadvantage of econometric-centric modelling is when the assumptions listed above fail, in particular super-exogeneity in the policy sphere Furthermore, it is reduced-form modelling in its nature, and even if super exogeneity holds, it may not be that structural relationships relating to individual agent preferences can be identified The advantage however is that it is not wedded to any particular economic theory, and by its agnosticism, can be a powerful tool for testing between economic theories On the other hand, economic-centric econometric modelling is inextricably wedded to the theory being proposed, and as such any estimation method is also constrained by this This can cause problems, not least from the possibility of omitted variable bias, meaning that econometric results are often not quite what the theorist is seeking for his or her theory Less formal or strict estimation methods are often employed in economic-centric modelling, on the basis that models are necessarily false and hence will be rejected by data These models nonetheless allow the counter-factual policy analysis described above, but their less strict estimation raises questions about their applicability: Although the policy analyses possible are attractive and insightful, it is possible they are based on false assumptions and hence are themselves false 7

8 For example, Ferreira-Lopes attempts to conclude about monetary union benefits based on monetary policy independence that does not necessarily exist, and ignores the imperfect nature of foreign exchange markets so powerfully alluded to by Buiter Thus, neither school of modelling can claim to be able to offer unambiguously better and more practical solutions, but each adds a different dimension to the discussion of the Swedish question thus far Buiter s lucid analysis is descriptive, and the work of Ferreira-Lopes and Söderström invoke economic theory in the form of DSGE models to investigate the problem We, along with Pesaran et al (2007), bring the tools of econometrics to the board Econometric methods, when appropriately employed, provide an important angle: that of what the data are revealing The model is able to indicate, but not prove, the existence of many assertions about the nature of the relationship between Sweden and the eurozone that are often taken on assumption, in particular the assumption of independence of policy As such we argue our paper provides a powerful contribution to the important and timely question of whether Sweden should adopt the euro III Methodology Recently, in Reade and Volz (2009a) and Reade and Volz (2009b) we investigate monetary policy independence using the cointegrated vector-autoregressive (VAR) methodology In these papers we consider the dependence economies in regions exhibit on both regionally dominant economies, and internationally influential economies Specifically, we investigate whether cointegrating vectors, or steady-state relationships, exist between economies, and consider which economies adjust, and which remain exogenous We find little evidence of monetary policy independence for countries other than the US, Japan and the eurozone (or Germany within Europe, pre-emu) In Reade and Volz (2009a) the ability of the cointegration framework to discriminate between cases of monetary policy dependence and independence was emphasised by using the counter-example of the UK While cointegrating relationships are found between all countries that later became members of EMU, with Germany being the dominant player, no cointegrating relationship was found between the UK and Germany despite the UK s involvement in the ERM As with our earlier papers, we conduct our analysis using the cointegrated VAR methodology to inves- 8

9 tigate the degree of monetary policy independence that the Sveriges Riksbank enjoys In particular, we consider a two-country pairing, that of Sweden and the eurozone This somewhat mimics our paper on pre-emu Europe, where we considered combinations of countries with Germany This pre-emu analysis builds on Edison and MacDonald (2003), and is similar to the earlier-mentioned Global VAR model of Pesaran et al (2007), albeit on a smaller, more focussed, scale The Global VAR, as its name suggests, builds a model linking industries and economies together in one entity In Pesaran et al a counterfactual exercise is employed to ask what would have happened had Sweden joined the euro in 1999 Our emphasis is on the question of whether Sweden should join the euro now, and not what might have happened had it joined at some point in the past, and hence we report estimations purely on historical data to investigate this The majority of empirical investigations of monetary policy independence model interest rates in isolation, even though classical interest rate parity conditions involve exchange rates One exception to this is Fratianni and von Hagen (1990), who add in domestic inflation and growth as exogenous variables into a VAR model of the first differences of interest rates when investigating German monetary dominance in the pre-euro EMS 11 We follow in this trend of modelling only interest rates because if a sensible and reasonably well-specified econometric model can be found for interest rates in isolation then the so-called sectoral general-to-specific property of cointegration applies This theory states that if the information set is enlarged (to include exchange rates) then the same cointegrating vectors found in the smaller system will also be found in the larger system, can be invoked to support the analysis 12 If a well-specified model is found, this is evidence suggesting that other important effects in a different context (such as purchasing power parity) have been successfully partialled out of the analysis, in that their effect, while important, will not affect the parameters of the case of interest The cointegrated VAR methodology builds on the cointegration framework of Engle and Granger (1987) for modelling non-stationary time series, extending it into multivariate models If two time series are individually non-stationary, and if some linear combination of them is stationary, then these two time series are said to be cointegrated From an economic standpoint it is theoretically implausible that interest rates 11 For a review of empirical studies in this field see Reade and Volz (2009b) 12 For more on this property, see Ericsson et al (1994) and Juselius (2007) 9

10 can be non-stationary because this would imply that the variance of interest rates would be increasing over time to infinity Nonetheless, worldwide interest rates since the early 1970s do not satisfy the assumption of a constant mean and variance Such analysis brings into question the important matter of sample length; a longer sample size will usually reveal more mean reversion, since the longer the time period, the more time for a series to revert to its true mean level Given that over sample sizes of over thirty years it is very difficult to maintain that interest rates have stationary means and variances, this leaves us with somewhat of a quandry: do we take the more appropriate statistical representation of interest rates, or do we revert to what economic theory implies? It is often the case that a statistical approximation of non-stationarity is more appropriate and less harmful than obeying an theoretical economic regularity This is particularly the case with non-stationary time series; failing to account for near non-stationarity (if indeed that is the best characterisation of interest rates) may result in inaccurate inference, and possibly spurious significance of results Thus we treat both interest rates as non-stationary 13 Furthermore, given the analysis of Hendry and Clements (2001), it seems very likely that even stationary cointegrating relationships may have shifted over a time period the length of that considered here, with different exchange rate regimes being employed at different points The existence of stationary cointegrating vectors in the multivariate framework of Johansen (1995) is based on the correlation between such linear combinations of the data levels and the more stationary first differences of the data If structural breaks have occurred then it is unlikely that any such correlation would exist Thus if we do find a significant correlation in the data, this is suggestive that despite all possible structural change and other impediments to econometric analysis, some long-run steady state relationship does exist in the data The multivariate approach also allows each series in the system to be modelled, hence avoids any a priori assumptions about exogeneity of variables, and direction of causality This is especially helpful in the context of monetary policy independence While it is theoretically implausible that the eurozone might rely on Sweden for interest rate setting, it is still preferable that we do not rule out this possibility at the outset; it emphasises the credibility of the modelling strategy if clear patterns such as this are upheld by 13 ADF unit root testing is unable to reject the null hypothesis of stationarity in both time series 10

11 the model Furthermore, the multivariate context allows a much richer modelling of the dynamics of the variables under consideration, and allows more than one cointegrating vector to exist 14 Turning to specifics, we consider a bivariate system with data vector X t, given by: X t = ( r r ) t (1) The domestic interest rate (here the Swedish one) is denoted r t while the foreign interest rate, here the eurozone rate, is denoted by r i These two series form a vector autoregression: K X t = Π 0 + Π 1 t + Π i X t i + u t, u t N ( 0, σ 2) (2) i=1 Here, X t is a p T data matrix, while Π i are p p coefficient matrices The Π 0 and Π 1 matrices refer to deterministic terms: a constant and a trend term, which are allowed to exist within this framework, and can be restricted to zero if insignificant If the data are non-stationary, so X t I(1), then in order for (2) to be balanced (given the stationarity assumption on u t ), it must be rearranged into equilibrium-correction form: K 1 X t = Π X t 1 + Γ i X t i + u t, (3) i=1 where X t 1 = (X t 1, 1, t), Π = (Π, Π 0, Π 1 ), Π = K i=1 Π i I, and Γ i = K j=i+1 Π j We have banded together the coefficients for the lagged regressors and the deterministic terms, for ease of exposition Further, if X t I(1), then given that u t I(0) and X t I(0) then Π must be of reduced rank for (3) to be balanced If Π is of reduced rank then there exist p r matrices α and β such that Π = αβ, and (3) becomes: 15 X t = α β K 1 X t 1 + Γ i X t k + u t, (4) k=1 where β = (β, β 0, β 1 ) and X t 1 = (X t 1, 1, t) The β X t 1 terms are cointegrating vectors, the stationary relationships between non-stationary variables, or steady-state relationships In the interest rate context, they are combinations of interest rates that individually are non-stationary, but together are stationary, with the cointegrating vector being an interest-rate parity relationship 14 Of course, given we model just two series in this context, this latter advantage is not particularly relevant 15 With the appropriate similar transformations of the constant and trend terms between the cointegrating space and the data differences, Π 0 = αβ 0 + γ 0 and Π 1 = αβ 1 + γ 1 Because a trend in first differences implies an implausible quadratic trend, we restrict γ 1 = 0 We also restrict the constant to only lie within the cointegrating space, hence γ 0 = 0 11

12 If the Π matrix is of rank one, so that one cointegrating vector exists, then β is of dimension 1 p and hence β is 1 p + 2 including the constant and trend, and we might write β X t 1 as: β X t 1 = ( ) β 2 β 3 β 0 β 1 r t r t 1 t = β 2r t + β 3 r t + β 0 + β 1 t (5) If we find rank one, we will likely find some relationship between the two interest rates, which is evidence of monetary policy dependence between the two economies The α coefficients also allow extra insight into the economic dynamics taking place in the data, as they dictate whether a variable adjusts to a particular cointegrating vector, and the speed of that adjustment, if any is found It may be expected that a large or dominant economy, such as the eurozone, would not adjust to this cointegrating vector, as it might be expected to exert monetary policy independence; so if we assume r to be the larger economy, then one would expect α 2 = 0 The smaller economy may be expected to adjust, so α 1 0 Furthermore, α 1 describes how much of any disequilibrium is corrected each period, as α = X t /(β X t 1 ), hence (ceteris paribus) a speed of adjustment can be calculated; the smaller is this coefficient, the more independent is a country in setting its monetary policy, as it devotes less of its attention to correcting to what other interest rates are doing As such, the α matrix is very informative about the nature of monetary policy independence An economy not adjusting to a cointegrating vector in which it appears is said to drive the system: the level of the interest rate in that economy is not constrained by the cointegrating relationship, but instead dictates what level that cointegrating relationship takes At this point we should emphasise that cointegrating vectors describe equilibrium combinations of economic variables, and hence there could be high interest rate equilibrium combinations as well as lower interest rate equilibrium combinations One potential concern might be that if we find a cointegrating vector, it must be that at least one of the two economies will adjust, otherwise the cointegrating vector would not be found (it is found by considering the eigenvalues, which are the squared correlations between the first differences X t and the linear combinations of levels, β X t ) situation where the eurozone adjusts Hence for Sweden to not adjust, we would require the implausible Given this, we would expect Swedish monetary policy independence 12

13 to be manifested in a lack of cointegration between the two series If it happens that they do move together as time series, then economics and common sense dictates it must be the eurozone influencing Sweden and not vice versa; thankfully also we can find the direction of this relationship out in the data without imposing economic structure a priori using the cointegrated VAR framework As will become evident in the next section on data selection, as well as in Section VI where we examine various other financial and real economic variables, many patterns of convergence exist between Sweden and the eurozone We hence add a time trend term into our model that is able to capture convergence In the face of convergence between Sweden and the eurozone it would be somewhat nonsensical to estimate a model which is unable to capture that convergence Although the model assumes that rate at which Sweden or the eurozone adjusts to any cointegrating relationship remains constant throughout the sample, the trend term allows any cointegrating vector to reflect a gradually tighter and tighter relationship between interest rates Naturally, such a model will be of marginal use for forecasting longer horizons, since a linear time trend at some point will begin to imply divergence, which is implausible However, the main intended purpose of this model is not to forecast but instead to understand more the historical co-dependence between Sweden and the eurozone block of European economies 16 IV Data From the policy interest rates of the ECB and the Sveriges Riksbank plotted in Figure 2, initially after 1999, the eurozone rate appears to lead the Swedish rate, and although that gap closes up to 2009 and the current global recession, it is clear the eurozone still leads the Swedish rate, simply the lag is less between the two For our estimations, we use daily interbank interest rates due to their relevance for monetary policymaking and their availability and variation at high frequency To the extent that markets factor in anticipated monetary policy movements before they happen, then interbank rates are a good reflection of monetary policy Furthermore, if the setting of the interest rate at which banks operate in a particular economy reflect the monetary needs of that economy, then that country can be said to have monetary policy independence If 16 Of course, this model could be used for forecasting: Simply we do not do that in this paper In terms of whether a model such as this should be used for forecasting, Hendry and Clements (2001) give a number of reasons why this model may be less appropriate, particularly over shorter forecast horizons 13

14 75 Eurozone Sweden I n te r e st 50 R at e Time Figure 2: Central bank policy rates (Riksbank repo rate, ECB refinancing rate) Source: Datastream instead the rate at which banks are able to lend is dictated by other influences, such as the economic events in other economies, then clearly this economy exhibits little if any independence in practical monetary policymaking Hence interbank interest rates provide an effect means for investigating monetary policy independence Our data are three-month interbank interest rates sampled at a daily frequency for Sweden (the Stockholm Interbank Offered Rate, STIBOR) and the eurozone (the European Interbank Offered Rate, EURIBOR) We begin estimation in 1987; pre-euro, the rate was simply calculated for banks in what became the euro area, and post 2001 Greece is included in this calculation The data begin on 1 January 1987 and our last observation is drawn on 19 June 2009, giving 5,794 observations Our dataset includes a number of periods of structural instability, notably the Swedish banking crisis of 1992, the structural break of the introduction of the euro in 1999, and the current financial crisis It is our belief that the credibility of econometric analysis is not enhanced by the judicial use of data via the selection of convenient sub-samples of the available sample Hence we estimate over the entire time period for which daily data are available for Sweden and the eurozone An important point here is that the euro only came into being part-way through our sample; one might thus be inclined to wonder whether it would make more sense for our analysis if we only considered the 14

15 post-euro portion of our data? With daily observations, we have many observations, and hence cutting the sample to just post-1999 would still provide well over a thousand observations A crucial distinction however must be made between enlarging a dataset by raising the frequency of observations, and by increasing the time frame We are considering movements between macroeconomic variables over the business cycle, and as such, when only considering 10 years of data, increasing the frequency will not produce more economic cycles in which to observe the co-movements between the data series 17 Cointegration, like stationarity, is better observed over longer time periods, particularly when macroeconomic variables are under scrutiny, and as such it is likely that over a shorter time period there are less corrections to the equilibrium in order for that equilibrium to be effectively discovered Reflecting this, we carried out the cointegration analysis included in this paper over just the post-1999 data period, and found results extremely similar to those reported in the paper; the only notable difference was that the cointegration discovered was slightly weaker The main reason for this can be found from considering the actual cointegrating vector found over the entire data period, plotted in Figure 7; over the longer horizon it is quite clear that the equilibrium relationship exists, as there are many corrections to it But if one takes a smaller segment of that relationship, such as that since 1999, then many of these equilibrium corrections are lost and the equilibrium relationship that can be found is weaker It should be emphasised that this is a time series statistical feature and not an economic feature: Cointegration is a stationarity concept and can only be found if these is sufficient data to display the reversions to mean Hence it will not necessarily be found as strongly over shorter time horizons, and as such taking both time periods in isolation (pre- and post-emu) results in weaker cointegration 18 The two series are plotted in Figure 3 There appears to be clear co-movement between the series, and arguably some convergence also In the late 1980s and early 1990s there is a marked difference between the two series, with Sweden s interest rates markedly higher throughout The 1992 ERM crisis and associated attempts to defend the value of the currency have a huge effect on Swedish interest rates Apart from a bulge 1996, post 1992 the two interest rates appear to move much more closely together, with periods in the 17 Although, ceteris paribus, it will increase the precision with which effects in the time period are measured 18 The highest eigenvalue, which can be taken as a measure of the strength of cointegration, in the entire sample is 0009, while it was 0013 pre-emu and 0006 post-emu Sweden s α coefficient, hence measuring Sweden s level of independence, was 0003 post-emu, 0009 pre-emu and 0007 overall, with all three coefficients strongly significant 15

16 I n te r e st R at e Eurozone Sweden Time Figure 3: Interbank interest rates for the eurozone and Sweden, 1/1/1987 6/19/2009 Source: Datastream early part of the 21st century where the Swedish interest rate is below that of the eurozone The effect of the recent financial crisis is evident in the sharp fall in interest rates at the end of 2008 We use data at a daily frequency which provides plenty of observations, but also potentially adds noise to our dataset, and will likely produce fat-tailed residual distributions due to ARCH effects An important question is whether there exists particularly much variation between observations at such a high frequency Figure 4 plots the interest rate series over four shorter time periods to give some idea of whether the plot in Figure 3 hides the variation (or lack of it) in the series because the scale is so large These four plots show that even in the more recent time periods there is considerable variation in the interest rate series, albeit on a smaller scale to earlier in the sample However, such heteroskedastic effects, while interesting and warranting future investigation, do not induce bias in coefficient estimates, nor do they affect the rank test outcome in the cointegrated VAR (Nielsen and Rahbek, 2000) V Results We estimate the model with six lags (days) of both variables, and we include 46 impulse dummy variables to cope with outliers The majority of these outliers are to cope with the 1992 ERM crisis, as can be seen in Figure 5, which plots the two interest rates with the dummy variables added plotted directly below, 16

17 Sweden Euro Figure 4: Interest rates broken into four smaller time periods (note different scales on vertical axes in each case) Source: Datastream corresponding to particular observations 19 Concern may be raised that by using dummy variables to cope with perhaps the most noteworthy historical event specific to our sample, we are choosing to omit, or ignore that information Dummy variables are used to model outlier events, events that plausibly belong to a different distribution to that of the rest of our data series Undoubtedly there is information in the financial crisis of note for our estimation purposes, and it should be emphasised that as our data is daily we by no means eliminate all observations corresponding to the crisis, and hence plausibly we retain all the relevant information for our purposes It is well known from the ARCH literature (eg Engle, 1995) that at times of volatility high frequency financial data series will display greater noise thus arguably some different 19 It should be emphasised here that the asymptotic impact of entering such dummies given the dynamics of the VAR system have been extensively studied by, amongst others, Johansen et al (2000) Furthermore, much research has been devoted to the issue of using impulse dummies for outliers in recent years (eg Hendry et al, 2008) and the main implications of that literature is that the inclusion of irrelevant dummies is harmless, but not including dummies for outliers can distort inference As such, we are confident about our use of such dummy variables 17

18 30 Sweden 25 Eurozone I n te r e st R at e D u m m ie s Figure 5: Plot of eurozone and Swedish interbank interest rates with impulse dummies also plotted Source: Datastream Time -1 distribution of very high variance and unknown mean, and we argue that we are removing this noise from our data as opposed to omitting all information relevant or not from this notable historical incident By omitting such noisy observations we are able to better uncover the true movements in the data series which might otherwise be dwarfed by the volatility in interest rates surrounding the crisis With the addition of these dummy variables, which produces a total number of parameters in our model of 124, a model is arrived at which satisfies the requirements for a model to produce an accurately sized rank test, and a model which represents the data effectively 20 The model exhibits non-normal residuals due to the fat-tailed characteristic of high-frequency financial data, as usually modelled using ARCH-type models Because ARCH effects do not adversely impact the rank test, and because the residuals in the model, despite being non-normal, are symmetric, we choose to proceed As a result of such high frequency data, we also note signs of autocorrelation in our residuals This autocorrelation was independent of lag length, hence we retained a lag length of six Additionally, the 20 By stating our model represents the data effectively we state that all statistical evidence points towards our model yielding an accurately sized rank test and unbiased parameter estimates Whether the data represent the underlying economic reality is another matter, but in the case of interest rates there is little scope for measurement error, and the correspondence between interbank interest rates and monetary policy rates has already been discussed 18

19 Rank Eigenvalue Log-Likelihood Test statistic P-value ** Table 1: Trace test outcome for various rank possibilities * denotes a test rejection at the 10% significance level, ** denotes a rejection at the 5% significance level Source: Datastream rank test outcome is very conclusive at rank 1, as can be ascertained from Table 1, and hence we are not worried about any possible problems with testing that this autocorrelation might induce (Juselius, 2007) It is only in marginal test decisions where size distortions in testing are present that decisions may be affected, but thankfully as our decision is clear in favour of rank one, we can be confident despite the autocorrelation in our residuals From Table 1, the null hypothesis of rank zero is heavily rejected with a test statistic of 5538, but the null hypothesis of rank 1, to which we proceed when following the Johansen (1995) procedure, is not rejected and has a p-value of 77%, suggesting a very clear conclusion in favour of a rank of one Many alternative rank testing procedures have been proposed to cope with poor size and power properties in the trace test, not least the Bartlett corrections of Johansen (2002) 21 However again given our test outcome is very conclusive, it is unlikely that any size distortions in the trace test could have affected the test enough to alter this particular implied outcome The restrictions imposed allow for a (1,-1) relationship between the eurozone and Sweden, and also restrict the eurozone s adjustment coefficient in the alpha matrix to zero The Swedish adjustment coefficient has a t-statistic of around 7, hence is very significant Thus the eurozone can be seen to be driving the system These over-identifying restrictions are tested using a likelihood ratio test, and are accepted with a test statistic of 123, which has a p-value of 542% The resulting system is, where r t is the domestic (Swedish) interest rate, and r t is the foreign (eurozone) interest rate: ( ) ( r r = (0001) 0 ) ( r r 2343 (0271) (000008) t ) (6) Sweden s adjustment coefficient is low at 0007, but it is of the right sign, and it should be emphasised that 21 The Bartlett corrections are most necessary in the case of I(2), or near-i(2) data As it is clear that our data are not I(2) from a brief consideration of the data plot (I(2) data tend to be very smooth, something which interbank interest rates are not), then we are confident in not using a Bartlett correction 19

20 we are considering daily data The half life of any deviation from equilibrium is about 100 days, hence about a third of a year The clear implication of this is that Sweden follows the eurozone in its interest rate setting Its monetary policy may have the illusion of independence due to standing outside the eurozone, yet its decisions are severely hampered by the need to follow the movements of interest rates in the eurozone The time trend term is also small, at 00006, but again we are considering daily data, and additionally it is of the right sign; if we consider that the cointegrating vector term is effectively equal to an I(0) error term, say ɛ t, then we can write: r r = 2343 (0271) (000008) t + ɛ t, (7) and hence the difference between the two interest rates is decreasing over time Multiplying the coefficient by the number of observations in our sample, 5,794, suggests that the difference between the two rates has shrunk by 348 basis points since 1987 This appears to tally with the divergence in interest rates of around three percentage points in the early part of the sample compared to the end, from Figure 3 The cointegrating vector itself is plotted in Figure 7, and it can be seen that apart from the drastic departure from equilibrium around the 1992 financial crisis, the vector is very close to equilibrium The vector is certainly stationary: it crosses the zero line on many occasions as a result of the long sample size, showing the value of using as many observations as possible As with macroeconomic movements, often the departures from equilibrium are quite sustained over economic cycles, and as such had data from only 1999 onwards been used, many equilibrium correction movements would have been omitted from the dataset, explaining the weaker cointegration result found over that reduced time period Estimating only over the post-euro period is a very simple recursive test; more detailed recursive analysis reveals minor instabilities around the 1992 financial crisis, also induced by the relatively few observations in the recursive sample at that point Estimating just the post-1992 period, the rank one test outcome remains strong, and although there are slight differences in the nature of the cointegrating relationship, the conclusion of Sweden s dependence on the eurozone remains Given the discussion of the need for a long sample in terms of time horizon and not frequency of observations, a recursive analysis is somewhat less useful here, given that the two variables display little evidence of structural breaks despite the change in exchange rate regime in 1992 Despite the 20

21 0 1 betares1 +/ 2SE betares2 +/ 2SE betares3 +/ 2SE Figure 6: Resursive plots of the three parameters in the cointegrating vector: The coefficient on the eurozone interest rate, on the constant and on the time trend noted disadvantages of recursive analysis in this context, it does provide an important contribution to our understanding of the Swedish situation As such, recursive estimates for the coefficient on the eurozone interest rate (which we successfully restrict to one in the reported analysis), the constant and the time trend are reported in Figure 6 There is a noticeable blip for the 1992 crisis, although part of the instability before 1992 is caused by a lack of observations rather than any inherent instability, necessarily Since the early-1990s the plots have been remarkably smooth and stable Even with the 1992 blip however it is clear that the level of the coefficients emerging prior to the crisis remains afterwards: The parameters in our model appear robust to this large and extraordinary economic even that took place Parameter stability from recursive analysis is used by Christensen and Nielsen (2009) to argue that their model is robust to the Lucas Critique, and in the terminology of our earlier discussion, their model satisfies super exogeneity We can thus apply the same test to our model and conclude that our parameters are invariant to extraordinary events such as the 1992 financial crisis, and hence relevant to use when considering economic policy analysis Considering robustness a little more, it is beneficial to check that our results are driven by the signals in the data, and not the noise daily data, while potentially introducing precision via more observations, may also 21

22 introduce more noise Financial markets are renown for their autoregressive conditional heteroskedasticity; volatility one day breeds volatility the next day Thus it makes some sense to also aggregate our data to the weekly and monthly frequencies If the results are the same then we can be confident that the estimates we have at the daily frequency represent the signals transmitted in the data, and not the noise present At both the weekly and monthly levels, the rank one hypothesis is again found to be most appropriate via testing, and the restrictions imposed on the daily model are accepted with a p-value of 735% at the weekly level, and 398% at the monthly level The constant coefficient remains similar in the cointegrating vector, while the time trend is affected by the data frequency, as is the rate at which the Swedish interbank rate corrects disequilibrium Both however are consistent with the daily model with the time trend coefficient again implying that the two interbank rates have converged from a difference of around 35 percentage points over the sample period, and the Swedish adjustment coefficients implying roughly similar lengths of times for which Swedish rates could depart from the equilibrium relationship of nearly a hundred days Reflecting on this outcome, it may be that Sweden and the eurozone are reacting to the same shocks, and hence our identification of policy dependence is over-stated Hence, it may be that while data and hence econometrics pick up a relationship, it actually reflects independent responses to the same shocks This may be so, but in reacting to a macroeconomic shock, the eurozone is systematically taking the lead, and Sweden following; this we feel is evidence suggestive of a lack of policy independence It is worth emphasising that the methodology employed here is able to identify countries that plausibly exhibit monetary policy independence; in Reade and Volz (2009a) the UK is found to display independence as a cointegrating relationship did not exist between it and Germany, pre-euro, while all of today s members of EMU exhibited a clear dependence on Germany s monetary policy We should also reiterate that it is the joint finding of a cointegrating relationship, adjustment by a dependent economy and non-adjustment by an independent economy that yields a monetary policy dependence conclusion Cointegration is to be expected due to the correlation between interest rates within Europe and other capitalist economies, but it is the adjustment to this cointegration that marks out the cointegrated VAR methodology as insightful in this context It would be more difficult to conclude in favour of Swedish policy dependence if the eurozone also adjusted to the 22